Wang, Yuxuan (University of Chinese Academy of Sciences), Yang, Yang (Institute of Automation, Chinese Academy of Sciences), Lei, Zhen (Institute of Automation, Chinese Academy of Sciences)

Transferable Lane Detection Via Adjacency Guided Instance Interaction and Label Reassignment

Scheduled for presentation during the Video Session "On-Demand Video Presentations" (VP-VP), Saturday, November 22, 2025, 08:00−18:00, On-Demand Platform

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on April 2, 2026

Abstract

Accurate lane instance detection is critical for autonomous driving but remains challenging under occlusions and complex lighting. A key difficulty lies in preserving consistent instance representations and supervision quality when lane appearances are degraded or partially missing. Real-world lanes often exhibit strong structural priors, with spatial correlations among adjacent instances, which are underutilized by existing methods. To address this, we propose the Adjacency-Guided Instance Interaction and Reassignment Module (AGIR-Module), a portable feature enhancement module that improves instance-level representation and supervision. AGIR enhances lane instance features by predicting an adjacency matrix that explicitly models inter-instance relationships. The adjacency matrix is incorporated into the self-attention mechanism as a learnable bias. This promotes stronger interactions among neighboring lanes and suppresses interference from unrelated ones. In addition, AGIR introduces an adjacency-guided label reassignment strategy to improve supervision under occlusion or matching ambiguity. Instead of rigid optimal matching, this strategy reallocates supervision to structurally correlated predictions, guided by adjacency cues and matching cost. Together, these components strengthen feature learning and ensure robust optimization under challenging scenarios. As AGIR operates entirely at the instance level, it can be seamlessly integrated into existing instance-based lane detection frameworks. Experiments on benchmark datasets demonstrate consistent performance improvements, validating the effectiveness and transferability of our approach across diverse architectures.